Research On Key Technology Of Emotional Semantic Analysis On Scene Images

With the rapid development of multimedia technology, Internet technologyand social networks, information resources that people can access are veryabundant. The application fields of images are gradually increasing and thenumber is growing at an alarming rate because they are visual and containmassive strong integrated information. But the inherent properties of imagessuch as the structure complexity, information diversity and temporalmultidimensional, have caused it becomes a research hot in academic fields toorganize and manage massive image data, realize effective and fast retrieval.Thus, emotion semantic analysis and retrieval technology for images comes intobeing, which integrates related knowledge in the field of artificial intelligence,computer vision, pattern recognition, psychology and database management,analyzes the high-level emotional semantic that image data contain, aimes atobtaining the inner emotional semantic information of images and establishingpractical image retrieval system. Therefore, research on the emotional semanticanalysis and efficient retrieval technology for images has a wide application prospect and practical value.This paper chooses scene images as the research object and research themethod of emotional semantic analysis and retrieval for all kinds of sceneimages in SUN Database. From the bedining of establishing emotion semanticaccess platform for scene images in the opening behavior experimentalenvironment, OCC emotion model is selected and modified, semantic fuzzyproblems are analyzed for scene images in semantic understanding, weight valueand threshold value of BP neural network are optimized using Particle SwarmOptimization (PSO) algorithm, strong predictor is built in order to predictemotional semantic classification of scene images through combining the outputresults of the15BP neural network by Adaboost algorithm, scene imageretrieval platform suitable for large data processing is built based on MapReduceparallel programming model. The paper systematically researches the method ofemotional semantic analysis and retrieval for scene images. The mainachievements and innovations are as follows:(1) Aiming at the semantic fuzzy problems when people understand images,the method of automatic semantic annotation for scene images based on fuzzytheory is proposed. This paper selects10emotion words of which the correlationis strong with emotion semantic of scene images in the OCC emotion model asthe basic emotional values, and defines3membership variables {very, neutral,hardly} to describe people’s emotional degree when understanding scene images.Color vision characteristics of scene images are extracted using the irregularblock method based on weight. Semantic feature mapping is realized using T-S fuzzy neural network. Emotional semantic automatic annotation for sceneimages is realized by the basic emotion value and the fuzzy membership degree.Experiments show that emotion understanding direction of different users for thesame scene image is consistent, but the depth degree of understanding isdifferent.(2) OCC emotion model is improved. A new OCC affective modeling isproposed taking into accout cognitive factors such as mood, personality. Thispaper uses dimension measurement (PAD) model to describe emotional factorsto obtain the mapping matrix of OCC emotional words and P, A, D value anduses five factor model (FFM) to describe people’s character. The character isdivided into five categories. And then we get the mapping matrix of characterand emotion. Finally the quantified OCC emotion model is obtained using twomapping matrix.(3) In order to improve prediction accuracy of emotional semantic categoryfor the unknown scene image, the forecasting algorithm of emotional semanticcategories for scene images is put forward based on Adaboost-PSO-BP neuralnetwork. This paper applies improved OCC emotional model to extract semanticfeature vectors and uses particle swarm optimization (PSO) algorithm tooptimize weight value and threshold value of BP neural network. Regarding theoptimized BP neural networks as weak predictors, Adaboost algorithm groupsoutput of15BP neural network predictors to build strong predictor, whicheffectively improves the prediction accuracy of emotional semantic categories ofscene images. (4) Aiming at large-scale scene images, a retrieval scheme for scene imagesbased on MapReduce parallel programming model is presented in order torealize efficient retrieval. Using parallel processing architecture of Hadoopcluster, parallel storage and retrieval methods for massive scene images aredesigned. Distributed Mean Shift algorithm is applied to feature clustering ofscene image under the Mahout environment to realize parallel clustering andmatching of features and improve the retrieval performance. Experiment resultsverify the efficiency of the proposed distributed parallel retrieval method fromthe aspects of retrieval accuracy, speedup and efficiency, data expansion rate.This paper launches the research around emotional semantic understandingof scene images. Discussion and research are carried out from the followingaspects: acquisition of emotional semantic data, automatic annotation, emotionalsemantic category prediction and efficient retrieval of large-scale scene imagedata. From a theoretical point of view, the research does the abstract, analysisand formal representation for emotional semantic of scene images. From apracticle point of view, the research builds an experimental platform to verifyand analyze, and provides a new thought and approach for emotional semanticunderstanding of all kinds of image data.